bug#410: 23.0.60; display-buffer regression

[martin rudalics]

> The recent reimplementation of display-buffer in lisp resulted in the
> following changed behavior, which I consider a regression:
>
> 1. emacs -Q
>
> 2. M-x calendar
>    ==> The window is vertically split, the *scratch* buffer above and
>    the Calendar buffer is below in a smaller window sized to fit.
>
> 3. M-: (pop-to-buffer (get-buffer "*Messages*"))
>    ==> Now the *Messages* buffer is above and the Calendar buffer is
>    still below, but the windows are evenly split, i.e. the Calendar
>    window is no longer sized to fit but is too big.  Prior to the
>    reimplementation of display-buffer doing this step did not change the
>    height of the Calendar window.
>
> The behavior in step 3 results from the invocation of
> window--even-window-heights in the last clause of the cond in
> display-buffer, which happens because the *Messages* buffer exits but is
> not displayed.  If the sexp in step 3 contained *scratch* instead of
> *Messages*, the height of the Calendar window would not have changed.

Thanks for noticing this.

> Looking at the pre-reimplementation source of Fdisplay_buffer, it looks
> like the window heights should get evened out just as in the lisp
> reimplementation; nevertheless, this is not what happens in the above
> recipe with the older code.

It's because in the Elisp code I evened window heights whenever they
were not equal.  The original code evened heights only when the height
of the window to display BUFFER-OR-NAME was less than that of the
selected window.  I've tried to restore the original behavior.  Please
try again.

I'd like to make the window code resilient in a way that applications
like calendar can naturally set `window-size-fixed'.  As a consequence,
`display-buffer' wouldn't change the calendar window's size even it were
larger than the new window.  Setting `window-size-fixed' can currently
result in unexpected behavior.  Fixing this is non-trivial.

Also I'd like to give `split-height-threshold' and `window-min-height'
buffer-local semantics and maybe add a `window-max-height' variable so
to do `fit-window-to-buffer' and `shrink-window-if-larger-than-buffer'
implicitly during each window configuration change.  `window-min-height'
equalling `window-max-height' for a particular buffer would then imply
`window-size-fixed' for that buffer.

Eventually, a user should be able to change window configurations
arbitrarily with the calendar window always keeping its original size
(unless there's no other way, obviously).

martin